Improving Distributed File System Performance in Virtual Machine Environments

Virtual machine (VM) systems have traditionally used virtual disks for file storage. Recently, there has been interest in using distributed file systems as a way to provide data storage to guest virtual machines, with the file server running on the same physical machine. Potential advantages include finegrained data sharing, data protection, versioning, and backup to multiple guests from one central place. Unfortunately, distributed file systems often perform substantially worse than local file systems because of high network overheads, even when the file server is on the same physical machine. This paper proposes two mechanisms to reduce communication overheads of inter-VM file system operations: Inter-VM metadata sharing and Virtual Remote Procedure Call (VRPC). Inter-VM metadata sharing enables clients to directly read file attributes from the server without an RPC exchange. VRPC follows standard RPC format but uses interVM shared memory to exchange data between a file server and its clients, which cuts out a lot of communication overhead. We implemented these two mechanisms on the Xen virtual machine platform and adapted NFS version 3 to take advantage of these two mechanisms. For an Apache build workload, NFS with Inter-VM metadata sharing and VPRCs was only 6.3% slower than the Ext3 file system, while standard NFS was 31.2% slower. These results suggest that Inter-VM metadata sharing and VRPCs significantly reduce communication overhead between virtual machines running on the same hardware.